Long Range Facial Image Acquisition and Quality · Long Range Facial Image Acquisition and Quality 5 body is an light source and such images could be collected in total darkness.

Long Range Facial Image Acquisition and Quality

Long Range Facial Image Acquisition andQuality

Terrance E. Boult1,2 and Walter Scheirer1.2

Abstract This chapter introduces issues in long range facial image acquisition andmeasures for image quality and their usage. Section 1, on image acquisition forface recognition discusses issues in lighting, sensor, lens, blur issues, which impactshort-range biometrics, but are more pronounced in long-range biometrics. Section2 introduces the design of controlled experiments for long range face, and whythey are needed. Section 3 introduces some of the weather and atmospheric effectsthat occur for long-range imaging, with numerous of examples. Section 4 addressesmeasurements of “system quality”, including image-quality measures and their usein prediction of face recognition algorithm. That section introduces the concept offailure prediction and techniques for analyzing different “quality” measures. Thesection ends with a discussion of post-recognition ”failure prediction” and its poten-tial role as a feedback mechanism in acquisition. Each section includes a collectionof open-ended questions to challenge the reader to think about the concepts moredeeply. For some of the questions we answer them after they are introduced; othersare left as an exercise for the reader.

1 Image Acquisition

Before any recognition can even be attempted, they system must acquire an imageof the subject with sufficient quality and resolution to detect and recognize the face.The issues examined in this section are the sensor-issues in lighting, image/sensorresolution issues, the field-of view, the depth of field, and effects of motion blur.

Vision and Security Technology Lab, University of Colorado at Colorado Springs, Colorado 80918,[email protected] · Securics Inc., Suite 200 1867 Austin Bluffs Parkway, Col-orado Springs CO 80918 [email protected]

Pre-print of chapter to appear in the book Handbook of Remote Biometrics: forSurveillance and Security. The original publication is available athttp://www.springerlink.com.

2 Terrance E. Boult and Walter Scheirer

1.1 In the beginning: Let There Be Light

To recognize a face one needs an image with visible features, which requires that wecollect an image with sufficient light levels and quality. Understanding the impact ofillumination variation, or normalizing to reduce it, is by far the most well studied ofthe issues associated with lighting and face-recognition [17, 1, 8, 10, 29, 5]. Whilethis type of work is very important, it is more focused on algorithms and not acqui-sition, and hence not covered in this chapter. This section will focus on illuminationaspects associated with acquisition, in particular, collecting and measuring light.

When working at close range in daylight conditions, the issue of sufficient light-ing is not a critical concern. However, as one starts looking at long-range face-basedrecognition, especially for 24 hour “surveillance”, assuring sufficient light level iscritical. Addressing this raises 2 unique issues: how to measure those light levels,and what sensors to use to collect in lower light and/or long range settings.

Long-range face needs very long focal lengths, often in the range 800-3200mm.Combining distance with the inherent limits on optics results in high F-Numbers lev-els. For example the Questar Ranger 3.5 which is a portable telescope used in longrange surveillance, provides 1275-3500mm focal lengths, but it comes at a cost oflight, with the 89mm(3.5inch) providing F13.2 at 1175mm and F35 at 3500mm.1

Recalling that each F-stop is a 50% loss of light, this telescope will measure inten-sity that is orders of magnitude smaller than that measured with a more traditionalF4 lens used for close/moderate range face recognition. This need for light is evenmore exasperated by the need for faster shutter speeds to avoid motion blur issuesthat will be described later in this chapter. Understanding the available lighting forlong-range settings thus is far more important than for standard face recognition.A question then is how to report light levels for long range experiments, especiallyfor low-light conditions. This is important not just for scientific experimentation,but for practical concerns if one wants to determine if conditions are sufficient for aparticular system to operate.

The most common measure for low-light imaging is in terms of lux. Lux is ameasure of illuminance (the accumulated light energy reaching a surface), and mea-sures how much light is in the scene. Given the lux reaching a surface, and thebi-directional reflectance function of the material/subject, one can estimate the lu-minous flux (the light leaving the surface in a particular direction). Luminous flux ismeasured in lumens. One can also compute the Luminous emittance which is the lu-minous flux per unit area emitted by a source. Luminous emittance, like luminance,is measured in lux. Given the Luminous flux one can use field-of-view of the lensand its F-stop to estimate the amount of light reaching the sensor from the targets.When the models are done right, this can be effective in predicting the light reachingthe camera and hence the response of the sensor.

Unfortunately, to use this approach for long-range low-light imaging there area number of difficulties. First, the reflectance function of the face varies consider-ably across the population. More significantly, the reflection is directional and is

1 http://www.company7.com/questar/surveillance/querange.html

Long Range Facial Image Acquisition and Quality 3

impacted significantly by self-shadowing, so measuring the scene irradiance with atraditional lux meter is not very effective without accounting for reflectance, shadingand shadowing which requires a detailed calculation after measurement, making itdifficult to use without advanced computer models. Finally, an issue especially im-portant for low light settings is that, even higher end hand-held light sensors areonly effective down to .01lux. These sensors use a light-to-voltage conversion thatmakes them good for bright scenes. But even though their accuracy is officially ratedat plus or minus .01lux, in practice, it is quite tenuous below .1lux. In many of ourfield experiments, the available lux sensors report underflow or zero (it is too darkfor them to operate). There are higher end NVIS lux meters, such as the ANV-410and TSP-410, but these are significantly more expensive and still have the issues ofnot providing sufficiently directional measurements to measure light that will reachthe sensor.

There is an alternative, which is to directly measure the light leaving the face inthe direction of the sensor: luminance. The candela per square meter ( cd

m2 ) is the SIunit of luminance; nit is a non-SI name also used for this unit. A candela is a lumenper steraidan (solid-angle), so a cd

m2 (nit) is equivalent to a lumenm2 sr), where as a lux is

a lumenm2 there is no simple conversion between lux and nits without using knowledge

of the view subtended by the source (face), which varies with distance. Luminanceis valuable because it quantities describe the “brightness” of the source and doesnot vary with distance, whereas illuminance in lux (the ”light” falling on a surface)must be manipulated to estimate how much light there is to measure. Putting itanother way, illuminance is a good measure to use when asking how well peopleor cameras can function anywhere within a dimly lit environment, but luminance isthe better measure to use for how well they can view a particular target (see [14]).The question then is how to effectively measure luminance for long-range face,especially if experimenting in low-light conditions.

To address these problems, and provide for a simple in-field measurement, wehave adapted a different type of measurement sensor. Using a sensor originally de-signed for “sky quality” measurements or “sky darkness” measurements, we have adevice that can operate at much lower light levels and can measure a narrow enoughFOV to capture just the data of the face. The sensor being used is the SQM-L2,based on the TAOS TSL237S sensor, which is a light to frequency converter. TheSQM-L has an added lens so that the Full Width Half Maximum (FWHM) of thesensor is ∼ 20◦. The sensitivity to a point source ∼ 19◦ off-axis is a factor of 10lower than on-axis and fall off faster beyond that. We will be experimenting withadding a component for further restriction of the field of view. The SQM-L sensorreports Magnitudes/arcsecond, which is an astronomical unit of measurement, butwhich is easily converted into cd

m2 . If we let s be the SQM-L value reported, thenLuminance cd

m2 = 108000 ×10−0.4∗s.We use the SQM-L for long-range face experiments by having the subject look

toward the camera, so it has appropriate lighting falling on the face, and them aim-ing the sensor on the center of their face while holding the sensor about 18 inches

2 http://unihedron.com/projects/darksky/


away. At this range a face subtends approximately 18◦, i.e. the sensor is measuringthe light leaving the face and little else (but some care has to be used if there aredistant lights behind the subject, and not shadow the face from any light sourceswith the hand/sensor). We call this measurement the “face luminance” and con-sider it the most useful overall lighting measurement for estimating performance ofa long-range face-system in low-light conditions. This is really a measurement ofluminance but it can be converted to luminous flux using the area of a face and thesolid angle subtended by a face from the target range, which is simple scaling.

Fig. 1 Example of low-light long-range EMCCD imagery. The measured scene illuminance for theleft image was .01lux, and illuminance was not measurable for the other two images. The measuredFace Luminance left to right, was 0.089 nits, .0768 nits and 0.015 nits respectively.

For example in figure 1, we have long range images in low-light conditions.The images were obtained at approximately 100m with an F5.6 Sigma 300mm-800mm. Capture occurs under star-light conditions 60, 90 and 120 minutes aftersunset, with a street light 100m off on the subject’s left. We prefer the face luminancemeasurement approach because works in the low-light setting where we want tooperate and it already accounts for the complex lighting/face shape interactions andis easily converted to a direct measure of the luminous flux heading in the directionof the camera. It is also very easy to “measure” in the field: hold the sensor, facethe direction of the camera and push the button on the sensor and hold (maybe upto 60 seconds if it’s really dark), then read the measurements from the unit’s LEDs.These measurements are more repeatable and reliable than using simple lux-basedestimations of overall illumination and then trying to convert it to lux at the sensor.

The second major issue impacting light levels for acquisition is the inherentimaging system sensitivity. This is significantly impacted by the sensor. Again, sincelong-range face is generally for surveillance, there is a general need to consider low-light conditions.

One approach often suggested for dealing with low-light settings is the use ofInfra-Red sensors for face recognition. For long wave IR (8-14 microns), the human


body is an light source and such images could be collected in total darkness. Whilethere has been some significant progress in the area of LWIR face (see [25, 23, 4,11]), we believe LWIR is too limited for long range face for several reasons. First,the need for long focal length lenses and high-resolution sensors for long-range face- the combination of which are simply not available for LWIR. The resolution issueswill be discussed in the next section. The second limitation of LWIR is that sincelong-range face is usually for non-cooperative subjects; LWIR requires specializedenrollment whereas visible recognition can use standard intelligence photos.

For comparison, consider Figure 2, which shows example images (close range)of 3 different types of sensors: a standard visible image, an intensified image and athermal image. This dataset can be obtained, for US researchers, from the author.An interesting open research question is the development of a LWIR recognitionsystem that can operate with visible image galleries. That is, with some intitial workin the area converting thermal into visible images addressed in [6].

Fig. 2 Left to right shows the same subject in a normally lit visible light camera, a low-lightintensified imagery and in LWIR thermal imagery. The Intensified imagery was obtained using anAmerican Eagle 603U which is a GenIII+ intensifier (specs are the same as PVS-14 commonlyused by the US Military). The intensified image was captured by a IQ-EYE smart camera with1280x1024 resolution. The thermal (LWIR) sensor is an NYTEK WEB-50 Micro-Bolometer, 8-14micron sensor, with images captured from the analog 640x480 video output. The visible imageswere captured from an IQ-EYE 1megapixel sensor. Images have faces with 80 pixels between eyes,which is the lower end of is expected for good recognition.

The alternative for low-light operation is to use some type of intensified imagery.There are a few alternatives within this group ranging from the very common tube-based intensifier optically coupled to a CCD sensor to an intensified CCD, to thecurrent generation of Electron Multiplying CCD. In our early work in low-lightwe used tube-based intensifiers coupled with a CDD. One disadvantage of this isthe blurring induced by the micro-charge plates of the intensifier and the visible“channel” artifacts, which have also been noted by other researchers (see Figure1 in [22]). Our more recent work in long-range face/surveillance [24] has movedto using EMCCD technology, based on a Salvador Imaging camera using the TC285 Chip. This provides 1004x1002 pixel images with 8µ pixels and an overallquantum efficiency of 65%. This sensor can operate from full sunlight down to star-light conditions. While the full details of our long-range low-light experiments are


beyond the scope of this chapter, Figure 1 provides some examples of cropped facedata collected with an 800mm Sigma F5.6 lens at more than 100m under very lowlight conditions. Except in the first of these photos, the naked eye camera operatorscould not even see the subject was there let alone recognize them. These imagesshow there is potential for long-range low-light face recognition using EMCCDtechnology.

1.2 Resolution: what does it mean and how much do we need?

In acquisition, presuming we have sufficient number of photons, the next most im-portant issue is the resolution of the target. While it is quite common to hear peopletalk about resolution in terms of number of pixels, it is more accurate to talk aboutthe effective resolution. One can formally define this using the Modulation TransferFunction (MTF) of the imaging system, which can account for both blur and con-trast loss. Under some simplifying assumptions one can decompose the MTF intothe product of the optical (lens) MTF, the sensor geometry MTF, and the diffusionMTF [9].

The ability of a lens to resolve detail is usually determined by the quality of thelens, though some very high end lenses and telescopes are diffraction limited. Theeffective aperture of the lens diffracts the light rays so a single point in space formsa diffraction pattern in the image, which is known as the Airy disk. If the systemis not diffraction limited, then other lens artifacts produce patch such that differentrays leaving a single scene point to not arrive at a single point in the image, givingrise to what is called the “circle of confusion”, even though it can be a far morecomplex shape. Ideally, the circle of confusion will be smaller than a sensor pixel.

Most MTF tables provided by lens manufacturers (see [3]), will show the MTFas a function of image position or distance from the center of the image. MTF val-ues above .6 are considered satisfactory, while some lenses such as the Canon EF400mm f2.8 IS USM, which we use for some of our long range experiments, havea circle of confusion of .035mm and MTF values above .9 over the whole field ofview. Even when extending with the Canon 2xII extender (making it it an 800mmF5.6 lens), the MTF is above .7 everywhere. In general, zoom lenses will have lowerMTF because of the more complex lens designs limit the optimization. (Note: youcan buy adaptes for C-mount to Canon lenses, with complete rs232 based controlof lens parameters such as focal distance, aperature, and stabalization parameters.These adapters are open air but because the increase the separation to adapt the35mm format to C-mount they may degrate the MTF).

It is important that when working with long-range biometric the lenses be match,or over qualified for the sensor choice. Modern high-quality lenses are multipleelement multi-coated designs optimized by the manufacturers for particular sensorchoices and with particular wavelengths in mind. If you can see vignetting, spatiallyvarying blur (when focused on a flat target) or color “fringe” artifacts, find a better


lens. It is also important to note that few lenses are optimized outside of the visiblerange, so be particularly careful in choices if working in the NIR range.

In the remainder presume that the optics are properly adapted to the sensor suchthat the overall MTF is not significantly limited by the optics, atmospheric or bymotion blur because if those are the limiting factors, it makes little sense to discusssensor “resolution”. In practice, the the important consideration is that the blur ofthe system is less than a pixel, otherwise the image can be effectively down-sampledby a factor of the blur, and not loose significant information. If you are doing long-range biometrics, the minimum is to measure your effective blur, or you can waste alot of time working on issues which are limited by blur. In short a large sensor/imagesize with a blurred image is not providing the resolution you might think.

Assuming good optics, resolution for long range face becomes a question of en-suring enough pixels on face to support recognition, and sufficiently above the min-imum needed for recognition to deal with the loss of resolution do to atmosphericturbulence. Formal models for atmospheric loss have been derived in the literature.See [27]. Diving into those models is beyond the chapter, but using such models canestimate an atmospheric blur level for long-range face, and expand the resolution re-quirements by an equivalent amount. We have routinely expanded the 60 pixel IPD,used for close range face recognition, to 80 pixel IPD for our 100m experiments.

Given the desired goal of 80 pixels between the eyes, and an average physical sizeof 4 of between 60mm and 72mm, combined with pixel size (for example, 8 micronfor visible spectrum sensors, 15 micron for LWIR) and size of the sensor (in numberof pixels), one can then estimate the focal length needed to produce an image withthe necessary spatial resolution on the subject. Deriving the formula is left as anexercise for the reader. In doing so, don’t forget to account for any “adaptors”, e.g.converting a 35mm camera lens into a C-mount is a change in format and back-focaldistance that impacts the effective focal length.

In reviewing Table 1, you should note that lenses for visible sensors up to1000mm are readily available and up to 3500mm available as special order via fields“telescopes”. Most intensified CCDs are only 640x480 and LWIR sensors are only320x240 (though there are exceptions for both, there are no 1280x1024 LWIR sen-sors). Long wave IR lenses up to 300mm are available and up to 1000mm is a specialorder (and massive).

1280x1024 640x480 320x240Range (m)

50 333 625 1250100 667 1250 2500150 1000 1875 3750200 1333 2500 5000250 1667 3125 6250300 2000 3750 7500

Table 1 Focal lengths needed to achieve 80 pixel average inter-pupil distance for different sensorsizes.


1.3 The working volume: Depth of Field and Field of view

The working volume, the region where the subject is in focus and within the filed ofview, is clearly important for the acquisition system design. Depth of field definesthe ranges around the focus distance where in subjects will be sharp focus. DOFincreases with decreasing lens aperture and decreases with focal length, so for long-range face it is a much more significant issue. The depth of field for a lens is notsymmetric, with different formula for the distance in front of the focus plane andbehind.

Formally we one can derive these as:

Front depth of field =d ·F ·a2

f 2 +d ·F ·a(1)

Rear depth of field =d ·F ·a2

f 2−d ·F ·a(2)

where f is the focal length , F is the F number, d is the diameter of the circle ofconfusion, and a is the subject distance from the first principal of the lens to subject.Note that if f 2 < d ·F ·a, the rear depth of field is considered infinite.

The important things to note here is that the depth of field decreases with thesquare of the focal length and for long focal length lenses can be quite short. Focusis further exasperated by the fact that for long-range face the optical axis usuallydoes not intersect a ground plane where the target will be, because they will bewalking well above the ground, and thus there is no way to easily pre-focus theimage. Fast auto-focus or having subjects “walk through” the DOF region are themost common choices.

While DOF is directly impacted by distance, the FOV of a lens is not. Ignoringblurring, the field of view necessary to maintain sufficient resolution for long-rangeface is actually the same as that needed for near-field “non-cooperative” subjects.The increased resolution requirements to account for atmospheric blur does changeit, but the change is effectively the same as requiring a larger inter-pupil distance inpixels.

The more significant difference is that there are many near-field face applicationspresuming cooperative subjects at effective choke-points to limit subject position-ing. With non-cooperative subjects, a larger field of view is needed to allow for sub-ject movement. This is especially acute in maritime biometrics where the subjects,and the sensor, may be moving with the waves.

The FOV is defined by the combination of the sensor resolution and the focallength. Presuming a focal length just sufficient for the minimum resolution, providesthe maximum FOV. Again, one can easily derive, via basic geometry, the FOV, theassociated pixel resolution on the sensor, and the effective physical size at the Focalpoint of the working volume. Example figures are shown in table 2, with the deriva-tion of the formula left to the reader. In deriving the table, we presumed an I’D of80 pixels and an overall head size of 160 pixels, and assured the head is within theframe. Figure 3 puts that data into perspective, and also shows how the FOV affects


“time of target” if one is using a stationary camera aimed at a choke point. It’s notjust that the larger sensor gives a larger FOV, the larger FOV translates into moreframes on target as they cross the larger FOV.

Sensor resolution 2048x1520 1280x1024 640x480 320x240Usable size in pixels 1888x1320 1120x824 480x280 160x40Usable physical FOV (in ft) 5.8’ x 4.3’ 3.6’ x 2.6’ 1.5’ x 0.9 ’ 0.5’ x .1’Allowed height variation 25in 15.6 in 5.4 in 0.6 in

Table 2 Usable Resolution, and size of FOV; the maximum size can reasonably be used for facerecognition. Conservative estimates are half the sizes/times shown.

Fig. 3 Example showing image sequences of a subject exiting a doorway and how the sensorresolution and FOV affect the effective number of frames where there is sufficient face data forpotential recognition. Note how 640x480 is just large enough for a head, and would not capturegood data for someone significantly taller or shorter. The 320x240 sensor is relatively useless.


1.4 Motion Artifacts

The last significant “sensor” issue to be discussed on acquisition is motion artifactsincluding motion blur. In any face-based system with non-controlled subjects theissue of subject motion must be addressed. This section addresses some of thosemotion artifact issues.

At first one might again presume that these issues are the same for long-rangeface as they are for any non-cooperative subject. That is, in part, correct. Howeverthe long focal lengths necessary for long-range face can mean even a slight vibra-tion in the sensor mounting can produce far more signficant results. The vibrationsnear field imaging on a basic camera mount on the wall might produce unnoticableinterlace artifacts, but the same vibration magnified by a 800mm lens might tearthe image appart and seem as here simple: Do not even think of using an interlacecamera for long-range face recognition.

Beyond interlace artifacts, there are two major artifacts which impact longrange face. While they also impact near-field face recognition, the fact that non-cooperative distance subjects can be moving faster, or that the long-range sensorcould be on a moving platform induces greater potential for these issues to be ob-jectionable.

The first of these issues is the well known motion blur. It can occur because ofplatform motion, including vibrations, as well as because of subject motion. Wehave found that for long-range face with walking subjects, most of the gait cyclewill have noticeable vertical motion blur with a significant reduction at the top ofthe stride. Figure 4, shows an example with both a clear face image and variousimages showing motion blur. These images were with an 800mm f5.6 lens with ashutter speed of 1/30 of a second. These types of issues are further exasperated ifattempts are made to use slower shutter speeds, or if the camera or subject are on amoving platform such as a ship.

Fig. 4 Example of motion blur. Subject is moving at a walking pace toward the EMCCD camera.Images are taken at approximately 100M from the camera at dusk. The top of the walking strideproduces minimal motion blur. (Scene has approximately .04 lux, yielding face lumens of 0.115nits.)


A less well known issue, which may not be obvious at first, arises with modernCMOS sensors that use a rolling shutter. Before we describe the issue, go study theimages in figure 5 and see if you can discern what it is. Note the images have a veryfast integration time speed, (1/10000 of a second), so if you thought it was motionblur, think again. And it’s not a depth of field issue either.

Fig. 5 Examples of rolling shutter artifact

There are two primary reasons to use a rolling shutter. First it saves one transistorper cell compared to a true “snapshot” shutter, and secondly it allows the integrationtime to be almost equal to the frame rate without significant buffering or fast readout circuits. The concept is quite simple, think of it as having two pointers to sensorpixels, both “rolling down” the sensor. One pointer is for readout of data, and theother is the the reset or erase operation. The time difference between a row’s eraseand next read defines the effective integration time (shutter speed). Each pixel sees(and accumulates) the light for the same exposure time (from the moment the erasepointer passes it till it is read out), but that happens at different times.

All this sounds good - a wider range of integration times at a lower cost. So whatis the problem? Looking back figure 5 again, we’ll give you a hint. The wall tothe subject’s left is a normal doorway, it’s a vertical edge and “straight”. Your cellphone camera is almost certainly a rolling shutter CMOS sensor - you can try someexperiments on your own see how significant the skew, warp or wobble can be.

The issue for rolling shutters is that even with a short integration time, the shutteris capturing data at different times for the top and bottom of the image. In the ex-ample the camera was subjected to horizontal motion fast enough that the top of theimage saw the wall in a different position than the middle or the bottom. Now askyourself what that would be doing to your face recognition algorithm, and you’llstart to appreciate the issue and probably think twice about rolling shutter sensors,even if they are the cost effective solution for getting multi-mega pixel arrays.


This first section has reviewed the early, and more static aspects of image ac-quisition for long-range face. The next section examines approaches for controlledexperiments for long-range evaluation, which is a necessary precursor before wecan get into the impacts of weather and atmosphere.

2 Photo-heads: Controlled Experiments in Long-Range Face

Even after the images are acquired the atmosphere and weather impacts can be crit-ical for long-range face acquisition. Studying them is a challenge as it is hard tocollect enough data under varying conditions. To address this we designed a spe-cialized experimental setup we call Photo-heads. The setup of the initial photo-headexperiment is shown in Figure 6, and example images in Figure 7. This “photo-head” data is unique in that it is a well-known set of 2D images (FERET) that weredisplayed on a special LCD and then re-imaged from approximately 94ft and 182ft.(We are currently implementing another photo-head setup at much greater distances,with 3D animated imagery). At these distances we needed a very long FOV lens, forwhich we used Phoenix 500mm zoom lenses (for 35mm cameras), with C-Mountadapters and Panasonic PAL cameras. The marine LCD was 800x600 resolutionwith 300NITS and a special anti-reflective coating. For display the FERET faceimages were scaled up for display. As one can see from the examples in Figure7, which are all from the same subject, the FERET data has a range of inter-pupildistances, poses and contrasts. This re-imaging model allows the system to controlpose/lighting and subjects so as to provide the repeatability needed to isolate the ef-fect of long-distance imaging and weather. As one can see, the collection producedimages sufficient for identification but with the types of issues, e.g. loss of contrastand variations in size, that one would expect in a realistic long-distance collection.All experiments herein used FaceIt (V4), the commercial face-recognition systemfrom Identix. This algorithm was one of the top performers in the National FaceRecognition vendor tests [16]. These tests were completed un the 2001-2004 timeframe.

This photo-head dataset is well suited formally study the issues to be encounteredin using biometrics for long-range “uncooperative” subjects in surveillance video.One of the most controlled variations is what we call “self-matching” the probe andthe gallery are based on the same image, except that the probe has been subject tothe long-range (re)imaging process, atmospheric distrubances and the weather. Theself-matching experiments are tightly controlled - they have exactly the same poseand subject lighting conditions. For initial testing we used a camera at approximately15 ft and the rank-one self-matching performance was over 99%, showing the re-imaging process and LCD are not a significant issue. We then moved to the realphoto-head collections. We generally ran each data set, which includes 1024 images,with 4 images of each subject, every 15 minuets, with collections over 4 months.The resulting 1.5TB of photo-head data was included in the DARPA HBASE, andsubsets of the data are available from the authors. With 4 images per subject we can


Fig. 6 The Photo-head experimental setup. Two cameras are positioned at two different distancesfrom a mounted weather-proof LCD display on a rooftop. Data capture occurred from dawn tilldusk. Experiments were conducted over two years, capturing weather for all seasonal conditions.

use the BRR technique [13] to estimate standard errors and statistical confidence.All our graphs include such error bars, though for clarity it is often show only forthe first plot point as it usually does not vary much as we change “rank” in the CMScurves.

Fig. 7 Example Photo-heads: four different views of the same gallery subject taken at 100ft. Mov-ing left to right: image taken at dawn, mid-morning, early afternoon, and evening.


3 In the Middle: Atmospheric and Weather

The obvious utility of a photo-head setup is the ability to capture outdoor condi-tions at all time of the year. Clearly, harsh weather conditions will have a significantimpact on recognition performance, but, even seemingly good conditions can haveunexpected impacts on recognition performance, depending on the interaction ofatmospherics. Figure 8 shows the visual impact of weather in three different con-ditions captured during the photo-head collection. Note the images are rotated fordisplay, the white on the left edge of the middle image is the snow building up onthe top of the display.

Fig. 8 From left to right: clear conditions, snow conditions, and rain conditions.

Fig. 9 CMC curves under various weather settings with self-matching.

The two graphs in Figure 9 show the impact of different weather conditions onface recognition. These are semi-log cumulative match curves with error bars fromBRR. The curve shows the recognition rate on the vertical axis and the log “rank”used to decide correct recognition on the horizontal axis. Rank-N recognition meansthe person was within the top N scores of the systems.

Two things should be apparent from these graphs. First, looking at rank-1 recog-nition (or even rank-3), off the shelf systems are not sufficient for these ranges, even


under the best of weather conditions and ideal pose/expression. Recall, these areself-matching experiments, only the imaging systems, atmospheric and weather arestoping it from being identical images for probe and gallery).

Second, and not suprising, the far camera, at approximately 182ft, was muchmore significantly impacted by the variations in weather. (The best weather rank-1recognition at 182ft was < 70%). While it is not show here, increasing wind evenmore significantly impacted the system, in part because at these ranges even a smalldeflection of the camera causes significant blur and may take the face out of thesensors field of view. (With these long FOV lenses, we needed 30” housings whichthat increase wind loading.) These graphs are computed over more than 20,000 im-ages and with the “controls” of the photo-head collections we know the images areidentical, thus the variations are not artifacts of individual errors, pose or expressionchanges. The techniques of [15] improved performance slightly, not statistically sig-nificantly, in large part because they dont address blur or geometric distortion, onlycontract and dynamic range.

There are also some intially suprising results within these curves. If you look atthe far camera results, you will see that light-rain and mist are statitically better than“clear” days. Can you generate a plausable hypothesize why “clear” days were notbetter? We controled for reported wind speed, so it is not that.

Fig. 10 On the left: variations over the time of day. On the right: recognition rank for various“quality” images

In addition to variations due to obvious weather effects, our experiments alsoshowed that there were variations due to time of day. Atmospherics, such as thermalwaves, can have a significant impact on recognition performance. Figure 11 showsthermal activity in difference images computed from a base frame and subsequentframes from a sequence timed over an entire day. The four images (two from thefar camera and two from the near) shown are from two successive captures onlya few minutes apart. Note the significant variation in the far camera between thetwo capture instances. Beyond atmospherics, rapid natural lighting changes, suchas when the sun is shining down on the scene, and then is quickly hidden by apassing cloud, can also impact the collection. The significant variation visible in the


first image is likely due to this effect. But what about the others. You can see fromthe “structure” of the differences that its not just a shifting of the image significantdifferences are up and to the left of edges in the upper left , but down and tot heright on the lower/right part of the image. The “difference” patterns are more likelocalized zooms, probably caused by atmospheric lensing from thermals.

The impact of atmospherics and natural lighting changes on the far camera’srecogntion rate is shown in Figure 10. These differences are statistically significant.Note that to reduce the impact of pose and lighting variations, these images are usingthe exact same image on the display as in the recognition database the only vari-ations between the probe and the gallery are those caused by the imaging system.Recall that indoors at 15ft, the performance on this type of data is nearly perfect.Even with this very strong constraint we see that at 182ft on a clear and low-windday, for Rank N recognition the performance of one of the best commercial algo-rithms of its day is below 65% with N < 4 and still below 80% recognition rate evenwhen N is 10. Again, these are averaged over hundreds of trials with 1024 imagesper trial, so this is not a sampling artifact.

Fig. 11 Difference images highlighting thermal activity and natural lighting changes for a se-quence of frames captured several minutes apart. The first and third images are produced by thefar camera at 200ft, and the second and fourth images are produced by the near camera at 100ft.

A first guess might be that the weather impacts the raw “image quality”, whichis determining the performance. We examined various measures of facial imagequality and (to our surprise), many of the errors had nothing to do with humanperceived or measured image quality. While better quality images generally did dobetter, it was not as strongly related as one might hope. The right half of Figure 10shows some examples of the recognition rank (i.e. where the imaged ended up whenprobes are sorted by match score) for a collection of images from a “same image”experiment. Rank and image quality in this set were inversely correlated. A detaileddiscussion of problematic issues with quality is presented in the next section.

Our research set off to find the causes of this unexpectedly poor performance.After considerable investigation we hypothesized the poor performance was due inlarge part to error in localization of the eyes. In [18] we presented an analysis ofthis theory. To definitively show the cause we added registration markers within


ourt photo-head data to allow us to transform the original eye coordinates to pro-vide eye-locations in the captured images. The graphs above show the recognitionperformance (with error bars) for the off-the-shelf FaceIt algorithms and when useforced FaceIt to use the correct eye positions. The results on both cameras where sta-tistically significant, and when the eyes are corrected the performance both far andnear cameras are similar. These results are, of course, highly optimistic because thedata for correction is artificial calibration points and secondly this is self-matching,with the same image as probes and gallery so the near perfect recognition is to beexpected. It is important to note that the “eye-locations” being discussed are notjust an question of where in the image the eyes appear but how that position relatedto where it should be in the image. In the “good quality” images of Figure 10, thecorrected eye position is not in the middle of the eye! Atmospheric turbulence andlensing effect can distort the face image to the point that to work properly the sys-tem needs to use a different eye position for its coordinate system and normalizationprocedures. Many of the computed eye locations were visibly off the eye, and theaverage difference between the computed and FaceIt eyes was 6 pixels.

4 In the end: Measuring Quality

As mentioned in the previous section, an obvious guess is that weather and atmo-spherics reduce the raw “image quality”, which is why they reduce performance. Inorder to study this potential impact, we formally defined quality and tried to studyits relation to performance. We experimented with multiple measures of “quality”,including blur and contrast in various ways. We eventually defined a blind signal tonosie ratio estimator for facial image quality, based on concepts from [28]. The con-cept is that statistical properties of edge images change with quality and have beenshowen to correlated with underlying signal to noise ratios. In our experiments, ourderived measure is, under general conditions, better correlated with recognition ratesthan the other quality measures examined.

To derive this measure, suppose the probability density function of an edge in-tensity image, ‖ ∇I ‖ is given by f ‖ ∇I ‖ (·) whcich is assume to have mean µ .The histogram of edge intensity image I can be modeled as a mixture of Rayleighprobability density functions, and that can be used to show that an estimate thesignal-to-noise ratio (SNR), is given by

QS =∫

∞

2µ

f ‖ ∇I ‖ (r)dr

It has been proven that the value of QS for a given noisy image is always smallerthan the value of QS for that image with less noise. Zhang and Blum also show it itcan estiamte blur and is overall correlated with Signal to Noise ratio.

Choosing a fixed sized window around the eyes (examples are shown in Figure12), we can define the Face SNR image quality as:


Fig. 12 Window around eyes for various images qualities.

Q′ = ∑edge above 2µ’s pixels∑edge pixels

'∫

∞

2µ

f ‖ ∇I ‖ (r)dr (3)

This Face SNR IQ estimage, the ration of number of pixels above twice the meanstrength to the total number of edge pixels, is easily calculated and can be shown tobe a good approximation to QS.

The results from this estimator are well correlated with recognition rate. We tookthe images classified them into 5 bins using Q′, and then examined the recognitionrate for each subset. Figure 13 shows the correlation between quality and recognitionrate, with overall correlations of 0.922 and 0.930 for two different galleries of photo-head data. Beyond the Face-SNR IQ estimate, we also performed experiments withmultiple levels of blur, contrast, and multi-metric fusion - none were better than theblind SNR estimate. While at first this might seem signifcant, you should recall wewere looking to understand/mitigate the impacts of atmospherics and wanted to usethe quality predict, on a per image basis, if an image was going to be successful forquality. Unfortunately, a stong correlation was not sufficient for a good predictor.

We concluded that “quality” is indeed found in the recognition performance, noton what we “like” to imagine in some preconceived concept of quality or even ourour blind SNR estimates. Interestingly, recent NIST studies [7] [2] on quality assess-ment come to this same conclusion. For the iris work in [7], three different qualityassessment algorithms lacked correlation in resulting recognition performance, indi-cating a lack of consensus on what image quality actually is. In the face recognitionwork [2], out of focus imagery was shown to produce better match scores.

We had already shown that on an individual image level both precieved and mea-sured quality could be inversly related to rank, but also showed that quality waspositively correlated with overall recognition scores. We are not alone in this obser-vation. We note that more recently [2] showed that, on a per instance basis, whatis visually of poor quality produced good recognition results. good, it was not suf-ficient for per-image predictor. Reflecting upon this issue of quality a bit deeper,we began to wonder how to predict if an image would be successful and also howcompare different measures of “quality” for face recognition.

The concept is using some measure of the system to predict if a particular inputimage will be (or is) successfully classified by the system. That is, we could thresh-old on quality and say any quality less than 2 will fail. With such a model we cancompare the usefulness of different image quality measures.

The question then becomes to measure use the effectiveness of each predictor.Since this is measuring system performance, this then suggests that for a comparisonof measures what is needed is some form of a Receiver Operator Characteristic


Fig. 13 In this plot, larger is better for quality. Correlations for blind SNR-based face image qualityto recognition rate are 0.922 and 0.930. Experiments were also performed with multiple levels ofblur, contrast, and multi-metric fusion. None were better than the blind SNR estimate.

(ROC) analysis on the prediction/classification performance. In [12] and [21] wedefine 4 cases that can be used as the basis of such a analysis. Let us define:

1. “False Accept”, when the prediction is that the recognition system will succeedbut the ground truth shows it will not. Type I error of the failure prediction andType I or Type II error of the recognition system.

2. “False Reject”, when the prediction is that the recognition system will fail but theground truth shows that it will be successful. Type II error of failure prediction.

3. “True Accept”, wherein the underlying recognition system and the predictionindicates that the match will be successful.

4. “True Reject”, when the prediction system predicts correctly that the system willfail. Type I or Type II error of the recognition system.

The two cases of most interest are Case 2 (system predicts they will not be recog-nized, but they are) and Case 1 (system predicts that they will be recognized but theyare not). From these two cases we can define the Failure Prediction False AcceptRate (FPFAR), and Failure Prediction Miss Detection Rate (FPMDR) (= 1-FPFRR(Failure Prediction False Reject Rate)) as:

FPFAR =|Case2|

|Case2|+ |Case3|(4)

FPMDR =|Case1|

|Case1|+ |Case4|(5)

With these definitions, the performance of the different reliability measures, andtheir induced classifier, can then be represented in a Failure Prediction ReceiverOperating Characteristic (FPROC) curve, of which an example is shown in fig-ure 14. Implicitly, various thresholds are points along the curve and as the qual-ity/performance threshold is varied, predictions of failure change the FPFAR and


FPMDR just as changing the threshold in a biometric verification system varies theFalse Accept Rate and the Miss Detect Rate (or False Reject Rate). High qualitydata, which usually matches better, will generally be toward the upper right, withlow failure prediction false alarms (and lower failures overall), but when good qual-ity data does fail it is harder to predict it so more are missed. Lowest quality data isusually toward the bottom right, with few missed failure predictions, but more falsepredictions, as poor quality more often results in marginal but correct matches.

The advantage of using the FPROC curve as opposed to simple CMC or ROCcurves with the data segmented by quality (or any other predictor variable) is twofold: First it allows for a more direct comparison of different measures on the samepopulation, or a the same quality measure on different sensors/groups. Second, seg-mentation of data to generated CMC/ROC curves inflates the measure since it meansthe quality i data is not interacting with quality j data. Furthermore, it is not practi-cal to compare measures or sensors when each one generates multiple ROC curves,especially if trying to compare multiple different “quality” measures. The FPROCevaluation approach allows us to vary the quality threshold over the gallery andsee how it impacts prediction, while still maintaining a mixed gallery of qualities.The FPROC curve requires an “evaluation” gallery, and depends on the underlyingrecognition system’s tuning, sensors, and decision making process.

Fig. 14 FPROC for 4 different image quality techniques on 12,000 images, compared with thepost-recognition Failure Analysis from Similarity Surface Theory (FASST) technique, with anwithout out image quality as a feature dimension.

The impact of switching approaches from a standard multiple CMC/ROC eval-uation of image quality to the FPROC representation is noted in figure 14, wherethree different image quality techniques and a simple image qualityfusion scheme


are plotted. The underlying data is 12,000 images obtained in varied weather con-ditions outdoors. As can be seen, while our Face SNR estimage out performs theother quality measures in prediction, none of the image quality techniques are verypowerful at predicting failure. Thus, while image quality is a well correlated withrecognition overall, it can fare poorly on a per image basis where significant pose,lighting, contrast, and compression are allowed. In essence, any unconstrained set-ting where data collection is taking place.

Early on in our “quality” analysis, we introduced a compelling alternative ap-proach [12], which was to learn to predict when a system fails and when it succeeds,and classify individual recognition instances using the learning as a basis. Based onthe decisions made by a machine learning classification system, a Failure PredictionReceiver Operator Characteristic Curve can be plotted, allowing the system operatorto vary a quality threshold in a meaningful way. Failure prediction analysis of thissort has been shown to be quite effective for single modalities [12], fusion acrosssensors for a single modality [26], and across different machine learning techniques[19] [21]. The FPROC quality prediction resuls of [12] are compared with basicimage quality predictions in figure 14 and are clearly signficantly better.

Since the early observation on image quality, we have continued to build thealternative approach in the form of post-recognition analysis of the recognition scoredistributions. We call this analysis Failure Analysis from Similarity Surface Theory.Let S be an n-dimensional similarity surface composed of k-dimensional feature datacomputed from similarity scores. The surface S can be parameterized by n differentcharacteristics and the features may be from matching data, non-matching data or amixed set of both.

Similarity Surface Theorem 4.1 For a recognition system, there exists a similaritysurface S, such that surface analysis around a hypothesized “match” can be used topredict failure of that hypothesis with high accuracy.

While the (empirical) similarity surface theorem 4.1 suggests that shape analysisshould predict failure, the details of the shapes and their potential for prediction areunknown functions of the data space. Because of the nature of biometric spaces, thesimilarity surface often contains features at multiple scales caused by matching withsub-clusters of related data (for example, multiple samples from the same individ-ual over time, from family members, or from people in similar demographic pop-ulations). What might be “peaked” in a low-noise system, where the inter-subjectvariations are small compared to intra-subject variations, might be flat in a systemwith significant inter-subject variations and a large population. These variations arefunctions of the underlying population, the biometric algorithms, and the collectionsystem. Thus, with theorem 4.1 as a basis, the system “learns” the appropriate sim-ilarity shape information for a particular system installation. We have applied theFASST technique to a variety of different data sets, with implementations utilizingdifferent learning techniques and underlying features generated from the recognitionscores [21, 19, 26, 12]. Even if we cannot get good predictors from just image facequality data, the “quality” of face data for recognition can be learned from the distri-bution of scores after matching. Further, we have demonstrated a multi-modal fusion


approach [20] for this sort of failure prediction, which is able to enhance recognitionperformance beyond the best performing multi-modal fusion algorithms.

5 Conclusions

In this chapter, we looked at the issues in image acquisition that must be consideredfor effective long-range, facial recognition. As we have seen, both obvious and verynon-obvious issues arise in all aspects of the image acquisition process. We dis-cussed working volume and resolution issues that designer must consider. Lightingis always a challenge for outdoor acquisition, and problems multiply in low-lightconditions. We have had good success measuring “face luminance” as opposed toscene lux or illuminance. Further, a mega-pixel EMCCD sensor with high resolutionhas provided us with images of sufficient quality and spatial resolution for standardface recognition, which overcomes many of the problems faced when LWIR sys-tems. In general, with today’s technology, cheaper components (low-resolution, in-terlaced sensors, rolling shutters, cheap lenses) will often hurt performance, in spiteof their bargain price tag.

Designing a long-range facial recognition system requires extensive testing forvalidation. Our photo-head setup provided much insight into the effects of weatherand atmospherics on long-range data acquisition. Not only did we learn of the im-pact on raw recognition scores, but also the limitations of image quality, which hasbeen a traditional indicator of performance. Our observations have led us to definea new paradigm for image assessment, based on post-recognition score analysis.We believe this post-recognition analysis is a critical component to enhance perfor-mance, along with proper equipment selection and system design.

6 Face Image Acquisition Exercises

1. Using the standard formulas for illumination and luminance (or irradiance andradiance, your choice), sketch out the steps needed to determin the amount oflight reacing the sensor for a face that is 100m away from the sensor. Using thisdetermine if a 2Mega Pixel Camera, with 11micron pixels, fitted with a cannonEOS 400mm f2.8 lens with a 2x adapter could operate at difference scene lightlevels.

2. Derive a formula for the necessary focal length for long range face recogntion,at distance d, using a sensor with 1280x1024 pixels each of p microns accross.State any assumptions you need to make along the way.

3. Derive a formula for the operational volume, both width of the FOV and theDepth of field, for long range face recogntion, at distance d, using a sensor with1280x1024 pixels each of p microns accross with a cannon EOS 400mm lenswith a 2x adapter set at maximum aperature.


4. For a camera at 200m from a subject of interest, who is exiting a door, whatis the necessary sensor size to have at least 3 seconds of video with sufficientresolution for face recognition (and temporal fusion). It is useful to ask for 3seconds to ensure some frames at the top of the gate where there is minimalmotion blur. State any assumptions you need to make along the way.

5. For a subject walking at normal speed, determine the shutter speed to ensure theirwalking does not produce more than a .5 pixel motion blur.

6. Consider the design of the Photohead experriment. List 4 limitations of the ex-perimental design and suggest alternative designs that overcome these limitations(while on a university/student budget :-) ).

7. In the definition of Face SNR IQ, we constrained it to a narrow region around theeyes and nose. Discuss the advantages and disadvantages of this windowing.

Acknowledgments

This work was supported in part by the DARPA HID program, ONR contract#N00014-00-1-0388, by NSF PFI Award # 0650251, DHS SBIR NBCHC080054,ONR STTR N00014-07-M-0421 and ONR MURI N00014-08-1-0638.

References

1. Adini, Y., Moses, Y. and Ullman, S.: Face Recognition: The Problem of Compensating forChanges in Illumination Direction. IEEE Trans. on Pattern Analysis and Machine Intelligence19 (1997), no. 7, 721-732.

2. Beveridge, R.: Face Recognition Vendor Test 2006 Experiment 4 Covariate Study. Presenta-tion at the NIST MBGC Kick-off Workshop (2008).

3. Canon: Optical Termoninology. The EF Lens Work III, Canon Inc., Lens Products Group,2006, 192-216.

4. Chen, X., Flynn, P.J. and Bowyer, K.W.: IR and Visible Light Face Recognition. ComputerImage and Vision Understanding 99 (2005), no. 3, 332-358.

5. Chen, T., Yin, W., Zhou, X., Comaniciu, D. and Huang T.: Total Variation Models for VariableLighting Face Recognition. IEEE Trans. Pattern Analysis Machine Intelligence 28 (2006), no.9, 1519-1524.

6. Dou, M.S., Zhang, C., Hao, P.W. and Li, J.: Converting Thermal Infrared Face Images intoNormal Gray-Level Images. The 2007 Asian Conference on Computer Vision, 2007, II: 722-732.

7. Flynn, P.: ICE Mining: Quality and Demographic Investigations of ICE 2006 PerformanceResults. Presentation at the NIST MBGC Kick-off Workshop (2008).

8. Georghiades, A.S., Kriegman, D.J. and Belhumeur, P.N.: Illumination Cones for Recognitionunder Variable Lighting: Faces. Proc. of 1998 IEEE Conf. on Computer Vision and PatternRecognition, 1998, 52-58.

9. Hoist, G.C.: CCD Array, Cameras, and Displays. SPIE Optical Engineering, 1996.10. Jacobs, D.W., Belhumeur, P.N. and Basri, R.: Comparing Images Under Variable Illumination.

Proc. of 1998 IEEE Conf. on Computer Vision and Pattern Recognition, 1998, 610-617.


11. Kong, S.G., Heo, J., Abidi, B.R., Paik, J.K. and Abidi, M.A.: Recent Advances in Visual andInfrared Face Recognition: A Review. Computer Vision and Image Understanding 97 (2005),no. 1, 103-135.

12. Li, W., Gao, X. and Boult, T.: Predicting Biometric System Failure. Proc. of the IEEE Con-ference on Computational Intelligence for Homeland Security and Personal Safety (CIHSPS2005), 2005.

13. Micheals, R. and Boult, T.: Efficient evaluation of classification and recognition systems.Proc. of 2001 IEEE Conf. on Computer Vision and Pattern Recognition, 2001, I: 50:57.

14. Marasco, P. and Task, H.: The Impact of Target Luminance and Radiance on Night VisionDevice Visual Performance Testing. Helmet- and Head-Mounted Displays VIII: Technologiesand Applications. Edited by Rash, Clarence E.; Reese, Colin E. Proceedings of the SPIE, 5079(2003), 174-183 .

15. Narasimhan, S. and Nayar, S.: Contrast Restoration of Weather Degraded Images. IEEETrans. on Pattern Analysis and Machine Intelligence 25 (2003), no. 6, 713-724.

16. Phillips, P.J., Grother, P., Micheals, R., Blackburn, D., Tabassi, E. and Bone, M.: Face Recog-nition Vendor Test 2002 (FRVT 2002). National Institute of Standards and Technology, NIS-TIR 6965, 2003.

17. Phillips, P.J. and Vardi, Y.: Efficient Illumination Normalization of Facial Images. PatternRecognition Letters 17 (1996), no. 8, 921-927.

18. Riopka, T. and Boult, T.: The Eyes Have It. ACM Biometrics Methods and ApplicationsWorkshop, 2003, 33-40.

19. Riopka, T. and Boult, T.: Classification Enhancement via Biometric Pattern Perturbation.IAPR Conference on Audio- and Video-based Biometric Person Authentication (SpringerLecture Notes in Computer Science) 3546 (2005), 850-859.

20. Scheirer, W. and Boult, T.: A Fusion Based Approach to Enhancing Multi-Modal BiometricRecognition System Failure and Overall Performance. In Proc. of the Second IEEE Confer-ence on Biometrics: Theory, Applications, and Systems, 2008.

21. Scheirer, W., Bendale, A. and Boult, T.: Predicting Biometric Facial Recognition Failure WithSimilarity Surfaces and Support Vector Machines. In Proc. of the IEEE Computer SocietyWorkshop on Biometrics, 2008.

22. Socolinsky, D., Wolff, L. and Lundberg, A.: Image Intensification for Low-Light Face Recog-nition. In Proc. of the IEEE Computer Society Workshop on Biometrics, 2006.

23. Socolinsky, D., Wolff, L., Neuheisel, J. and Eveland, C.: Illumination Invariant Face Recog-nition Using Thermal Infrared Imagery. Proc. of 2001 IEEE Conf. on Computer Vision andPattern Recognition, 2001, I: 527:534.

24. Vogelsong, T., Boult, T., Gardner, D., Woodworth, R., Johnson, R. C. and Heflin, B.: 24/7Security System: 60-FPS Color EMCCD Camera With Integral Human Recognition. Sensors,and Command, Control, Communications, and Intelligence (C3I) Technologies for HomelandSecurity and Homeland Defense VI. Edited by Carapezza, Edward M. Proceedings of theSPIE 6538 (2007), 65381S.

25. Wilder, J. and Phillips, P.J. and Jiang, C. and Wiener, S.: Comparison of Visible and Infra-red Imagery for Face Recognition. Proc. of the IEEE Conf. on Automated Face and GestureRecognition, 1996, 182-187.

26. Xie, B., Boult, T., Ramesh, V. and Zhu, Y.: Multi-Camera Face Recognition by Reliability-Based Selection. Proc. of the IEEE Conference on Computational Intelligence for HomelandSecurity and Personal Safety (CIHSPS 2006), 2006.

27. Yitzhaky, Y., Dror, I. and Kopeika, N.: Restoration of Atmospherically Blurred Images Ac-cording to Weather Predicted Atmospheric Modulation Transfer Function (MTF). OpticalEngineering 36 (1997), no. 11.

28. Zhang, Z. and Blum, R.: On Estimating the Quality of Noisy Images. Proc. of the IEEEInternational Conference on Acoustics, Speech, and Signal Processing, 1998, 2897-2900.

29. Zhao, W. and Chellappa, R.: Illumination-Insensitive Face Recognition using SymmetricShape-from-Shading. Proc. of 2000 IEEE Conf. on Computer Vision and Pattern Recogni-tion, 2000, I: 286-293.

Long Range Facial Image Acquisition and Quality · Long Range Facial Image Acquisition and Quality 5 body is an light source and such images could be collected in total darkness.

Documents